NVIDIA unveiled the Vera CPU on March 16, 2026 during GTC 2026, introducing what the company describes as the world's first processor purpose-built for the age of agentic AI and reinforcement learning. The announcement represents a fundamental architectural shift, with CEO Jensen Huang stating, "The CPU is no longer simply supporting the model; it's driving it."
88 Custom Cores Deliver 2x Energy Efficiency
Vera features 88 custom NVIDIA-designed Olympus cores with dual-task capability per core using NVIDIA Spatial Multithreading. The processor incorporates a second-generation NVIDIA Scalable Coherency Fabric and LPDDR5X memory subsystem, delivering up to 1.2 TB/s bandwidth—twice that of general-purpose CPUs at half the power consumption. Memory power consumption stays below 50 watts.
When paired with NVIDIA GPUs, Vera achieves 1.8 TB/s coherent bandwidth via NVLink-C2C, representing 7x the throughput of PCIe Gen 6. The processor marks the first CPU to support FP8 precision while maintaining full Armv9.2 compatibility. According to NVIDIA, Vera operates 50% faster than traditional rack-scale CPUs with 2x energy efficiency.
Real-World Performance in Streaming and Agentic Workloads
Redpanda testing demonstrated 5.5x lower latency on Apache Kafka-compatible workloads compared to traditional processors. A single rack configuration supports 256 liquid-cooled Vera CPUs, sustaining over 22,500 concurrent CPU environments at full performance—critical for large-scale agentic AI deployments.
The processor targets specific use cases including agentic AI and reinforcement learning, data processing and analytics, AI coding assistants and agents (with Cursor specifically mentioned), orchestration services, high-performance computing, and real-time streaming workloads.
Widespread Industry Adoption Across Hyperscalers and National Labs
Vera has secured adoption from major hyperscalers including Alibaba, ByteDance, Meta, and Oracle Cloud Infrastructure. Cloud providers CoreWeave, Lambda, Nebius, Nscale, Cloudflare, Crusoe, Together.AI, and Vultr are deploying the processor. System manufacturers Dell Technologies, HPE, Lenovo, Supermicro, ASUS, Compal, Foxconn, GIGABYTE, Pegatron, QCT, Wistron, and Wiwynn will offer Vera-based systems.
Software companies Cursor and Redpanda are integrating Vera for AI coding and streaming platform workloads. National laboratories including Leibniz Supercomputing Centre, Los Alamos, Lawrence Berkeley, and TACC are adopting the processor for research computing.
Vera is in full production with availability from partners beginning in H2 2026. The announcement gained 121 points on Hacker News with 76 comments.
Key Takeaways
- NVIDIA's Vera CPU features 88 custom Olympus cores with dual-task capability, delivering 50% faster performance than traditional rack-scale CPUs at 2x energy efficiency
- The processor achieves 1.2 TB/s memory bandwidth and 1.8 TB/s coherent bandwidth when paired with NVIDIA GPUs via NVLink-C2C—7x faster than PCIe Gen 6
- Vera is the first CPU to support FP8 precision while maintaining full Armv9.2 compatibility
- Major adopters include Alibaba, ByteDance, Meta, Oracle Cloud Infrastructure, and national laboratories including Los Alamos and Lawrence Berkeley
- A single rack supports 256 liquid-cooled Vera CPUs sustaining over 22,500 concurrent CPU environments at full performance